Random Forest Model Summaries
Joey Hulbert
2/18/2022
Purpose
The purpose of this markdown document is to list the steps we followed for refining the models.
Notes
- Aspects to consider or check
- There may be another set of soil variables that would be valuable to include
- There may be an updated set of normals to work with
- Notes from MC 2/18/22
- Steps
- 1 establish credibility
- establish approach is a robust technique and credible by sharing best model first
- First layout model - healthy and unhealthy - can see difference between healthy and unhealthy
- Able to say with certainty where healthy and unhealthy trees are located on landscape based on these variables
- establish approach is a robust technique and credible by sharing best model first
- 2 then explore other variables
- unpack model - increases uncertainty, but more interesting
- interesting to explore health characteristics and the variables that distinguish where they’re located on landscape
- compare thinning, dead top, healthy
- unpack model - increases uncertainty, but more interesting
- 1 establish credibility
- Steps
- Notes from writing paper (4.7.2022)
- How should we explain the steps toward selecting a model.
- Two primary analyses
- Climate data only with entire dataset
- climate and soils data with dataset limited to Oregon and Washington
- Both analyses followed some intiial exploration to simplify the number of groupings of environmental data.
- Two primary analyses
- How should we explain the steps toward selecting a model.
Data
Response variables
- Response variables to consider
- Number of unhealthy trees
- Tree canopy symptoms
Below we compare models using canopy symptoms as the response variable
Explanatory variables
- Explanatory variables included
- iNat data
- explanatory variables such as tree size
- Climate data
- 30yr normals 1991-2020
- last decade 2011-2020
- Soils data
- gSSURGGO
- muggaggat
- component
- gSSURGGO
- iNat data
Lets add summaries for how many variables each of these datasets provided
Note there may be a third soils dataset to incorporate. Also, need to confirm the normals data is actually the latest normal data.
The data used in the below models are described in the Data Wrangle folder.
Data wrangling
There are multiple methods to group the response variables deepening on desired resolution or fineness of the model.
- Response grouping options
- Original (no change)
- Filtered Symptoms
- We could drop trees outside of main symptom classes
- Note loss of data outside of categories
- We could drop trees outside of main symptom classes
- Reclassified symptoms
- Here we can created an ‘other’ variable to represent trees with the below symptoms
- extra cone crop
- yellowing canopy
- We can also group trees with ‘old dead tops’ and ‘new dead tops’ into a ‘dead top’ category
- Here we can created an ‘other’ variable to represent trees with the below symptoms
- Binary (healthy/unhealthy)
For now, we can move forward with the binary response grouping because it is the broadest and easiest for the model to classify with.
Filter Data
Filter trees to only those with soils data (Oregon and Washington)
All tree health categories
## # A tibble: 11 × 2
## # Groups: field.tree.canopy.symptoms [11]
## field.tree.canopy.symptoms n
## <fct> <int>
## 1 Branch Dieback or 'Flagging' 19
## 2 Browning Canopy 19
## 3 Extra Cone Crop 2
## 4 Healthy 403
## 5 Multiple Symptoms (please list in Notes) 17
## 6 New Dead Top (red or brown needles still attached) 33
## 7 Old Dead Top (needles already gone) 83
## 8 Other (please describe in Notes) 8
## 9 Thinning Canopy 118
## 10 Tree is dead 37
## 11 Yellowing Canopy 10We also need to filter the data to only include response and explanatory variables we’re interested in. For example, whether a sound clip was included in the iNat data is not important.
We also need to remove other response variables like “field.percent.canopy.affected….” so it is not used as a predictor for tree health.
Note it might be interesting to know if the user was an important factor in predicting if the tree is healthy/unhealthy.
There are also a number of factors that should probably be removed because they may be biasing the data. For example, only trees with the ‘other factor’ question may only be answered for unhealthy trees. We need to think about this a bit more.
Remove variables with variables that have near zero standard deviations (entire column is same value)
Impute data
Impute data
Overall Imputed data
We continue to get the below error, but were able to work around it by imputing the data.
Error in randomForest.default(m, y, …) : Need at least two classes to do classification.
To impute the data we have to remove factors with >53 levels.
The below code lists the number of levels for the variables that are factors.
- The following factors had more than 53 levels
- “muaggatt_musym”
- “muaggatt_muname”
- “component_compname”
- “component_geomdesc”
- “component_taxclname”
- “component_taxsubgrp”
Imputed data table
## ntree OOB 1 2 3 4 5 6 7 8 9 10 11
## 300: 46.06% 94.74% 94.74%100.00% 16.87% 76.47% 87.88% 71.08%100.00% 72.88% 91.89%100.00%
## ntree OOB 1 2 3 4 5 6 7 8 9 10 11
## 300: 45.79% 94.74% 94.74%100.00% 15.63% 82.35% 87.88% 72.29%100.00% 72.03% 97.30%100.00%
## ntree OOB 1 2 3 4 5 6 7 8 9 10 11
## 300: 45.39% 94.74%100.00%100.00% 14.39% 76.47% 87.88% 74.70%100.00% 72.88% 94.59%100.00%
## ntree OOB 1 2 3 4 5 6 7 8 9 10 11
## 300: 46.73% 94.74% 94.74%100.00% 16.38% 76.47% 87.88% 75.90%100.00% 74.58% 94.59%100.00%
## ntree OOB 1 2 3 4 5 6 7 8 9 10 11
## 300: 45.93%100.00% 94.74%100.00% 15.14% 82.35% 87.88% 75.90%100.00% 72.03% 94.59%100.00%
## ntree OOB 1 2 3 4 5 6 7 8 9 10 11
## 300: 46.06% 94.74% 94.74%100.00% 16.13% 76.47% 87.88% 73.49%100.00% 72.03% 97.30%100.00%Group response variables to binary values
Binary tree health categories
## # A tibble: 2 × 2
## # Groups: field.tree.canopy.symptoms [2]
## field.tree.canopy.symptoms n
## <fct> <int>
## 1 Healthy 403
## 2 Unhealthy 346Approach
- General approach
- build full model using binary response grouping variable
- binary response
- refine model to remove redundant explanatory variables
- build full model using binary response grouping variable
- Approach notes
- We need to train/test models to calculate correctly predicted presences (inverse of commission errors) and correctly predicted absences (inverse of omission errors)
- we may want to evaluate the climate parameters seperately
- we lose >300 observations when we filter out data to only include those with soil
- what happens if we build a model without soils data (without filtering to WA and OR)
Models
- Random forest model criteria
- nearzero variation columns were removed
- Data were imputed prior to model building.
- number of splits was not defined to allow automatic selection,
- ntree = 2001, importance = TRUE, proximity = TRUE, na.action = na.omit
- seed was set for reproducability with: set.seed(71)
Comparing full and reduced datasets of explanatory variables
- Datasets compared below
- Full model = all climate variables retained.
- Monthless model = filtered out the climate parameters for individual months
- Normal only, monthless model = removed decadal data
Full Model
##
## Call:
## randomForest(formula = field.tree.canopy.symptoms ~ ., data = binary, ntree = 2001, importance = TRUE, proximity = TRUE, na.action = na.omit)
## Type of random forest: classification
## Number of trees: 2001
## No. of variables tried at each split: 23
##
## OOB estimate of error rate: 26.44%
## Confusion matrix:
## Healthy Unhealthy class.error
## Healthy 303 100 0.248139
## Unhealthy 98 248 0.283237
Monthless Model
##
## Call:
## randomForest(formula = field.tree.canopy.symptoms ~ ., data = monthless.binary, ntree = 2001, importance = TRUE, proximity = TRUE, na.action = na.omit)
## Type of random forest: classification
## Number of trees: 2001
## No. of variables tried at each split: 15
##
## OOB estimate of error rate: 28.7%
## Confusion matrix:
## Healthy Unhealthy class.error
## Healthy 293 110 0.2729529
## Unhealthy 105 241 0.3034682
Normal, Monthless Model
##
## Call:
## randomForest(formula = field.tree.canopy.symptoms ~ ., data = normal.monthless.binary, ntree = 2001, importance = TRUE, proximity = TRUE, na.action = na.omit)
## Type of random forest: classification
## Number of trees: 2001
## No. of variables tried at each split: 13
##
## OOB estimate of error rate: 29.51%
## Confusion matrix:
## Healthy Unhealthy class.error
## Healthy 284 119 0.2952854
## Unhealthy 102 244 0.2947977
